Survival Is a Floor, Not a Forecast of Greatness
This chapter asks why early-stage venture screening can become better at identifying continued operation without becoming good at predicting exceptional growth. The subject is still the new venture facing an uncertain future, but the focus shifts from the content of a customer signal to the outcome that signal is allowed to predict. The distinction supports a practical judgment: a screening measure may help identify ventures less likely to disappear while saying little about which survivors will become winners.
The companion study, Predicting New Venture Survival and Growth: Does the Fog Lift?, gives a useful way to think about this separation. It follows 6,579 UK ventures that began trading in the same quarter of 2004 and tracks them for up to ten years. Its central question is whether the “fog” surrounding performance becomes thinner as a venture ages. The answer depends on which performance we mean. For sales growth, the fog remains dense. For survival, it clears somewhat.
Growth follows shocks; survival depends on a buffer
Imagine a new venture’s sales as a path that moves up and down in response to shocks: a customer cancels, a product works better than expected, a supplier fails, a regulation changes, or a founder finds a promising channel. In the study’s theoretical model, these changes are treated as broadly random from the standpoint of a prediction made in advance. The model does not claim that founders are inactive or that effort has no value. It says that deliberate actions occur inside a competitive environment where many outcomes remain difficult to forecast systematically.
That is what a random-walk description means here. The venture’s next change in size is not simply a continuation of its previous change. Past sales can tell us where the venture is, but they do not guarantee the direction or magnitude of the next movement. A venture that grew last year may decline this year; another that appeared stagnant may receive a contract that changes its position. The model therefore leaves much of the detailed path of sales unexplained. It is a disciplined statement of uncertainty, not a theory that nothing matters.
Survival is modeled differently. A venture begins with some stock of resources—financial capacity is the main intuition—and then experiences the same kind of shocks. It exits when its effective stock falls below a minimum threshold. The threshold need not be zero. A founder may stop before every resource is exhausted, or creditors may force closure while some assets remain. What matters is that the venture needs a buffer large enough to absorb adverse events.
This creates two distinct questions. Growth asks, “How far and how fast will the venture’s sales move?” Survival asks, “Will the venture remain above the point at which it can continue operating?” The first can remain volatile even when the second becomes more legible. A company can have unpredictable sales changes while still being well buffered against immediate exit. Conversely, a small venture can show an encouraging early sale and remain vulnerable because one setback would exhaust its limited resources.
The distinction also changes what an early screening signal could reasonably do. If customer transaction maturity is related to survival, it might indicate that a venture has begun to create market contact, revenue, or resource access that makes early exit less likely. It would not follow that the same measure forecasts the size of future sales jumps. A signal about the distance from an exit threshold is not automatically a signal about the height of the eventual peak.
The study’s framework can be represented as a sequence of states rather than as a simple upward growth curve:
The state transitions are the important part. A shock can push a lightly resourced venture across the exit threshold, while a buffered venture absorbs the same kind of disturbance and continues. The model does not say that every real exit follows exactly this path. The authors present it as a theoretical account of why survival may become more predictable even when growth remains close to a random process.
What ten years of evidence can—and cannot—show
The empirical results follow the model’s distinction. In the baseline growth regressions, the reported Nagelkerke R²—the proportion-like measure used to summarize how much the model explains—falls from roughly 27–37% in year 2 to roughly 13–22% by year 10. The exact range depends on the regression specification. In other words, the available variables become less useful for explaining the year-to-year sales-growth outcome as the cohort ages.
That decline is not simply evidence that all surviving ventures become more mysterious internally. When the researchers restrict the analysis to ventures known to survive the full ten years, the ability to explain growth remains roughly constant over time. This suggests that part of the declining fit in the full sample comes from selection: short-lived firms are easier to distinguish through their deteriorating or unusual account behavior, and their exit changes who remains in the later-year sample. Among the long-lived firms, growth remains difficult to explain rather than becoming steadily more predictable.
The R² figures also need the right interpretation. They summarize how much variation a model accounts for in a particular regression; they are not the probability that any one venture will survive, and they are not a ranking of the best companies. A lower growth R² does not mean that growth is meaningless or that every prediction is equally bad. It means that the variables in that specification explain less of the observed differences in annual growth as the cohort ages. Likewise, a rising survival R² does not mean that a decision-maker has found a complete theory of exit. It means that the included information separates continuing and exiting ventures somewhat more clearly in the later cross-sections.
The year-by-year design matters because the question changes with the venture’s age. A prediction made shortly after entry has little operating history to work with and must distinguish firms that have barely begun from firms that already have a resource cushion. A prediction in year ten is made after many firms have exited and after survivors have generated a record of endurance. Comparing the fit of those regressions is therefore partly a comparison of information environments. The later model can appear clearer not only because the venture has changed, but because the population has been filtered by what happened earlier.
Consider two ventures with equally uncertain sales paths. One begins with a thin buffer and depends on one or two early relationships; a single adverse shock can push it below the point at which it can continue. The other has more resources and several ways to absorb a setback. Neither path guarantees strong growth. The second may remain open while its sales fluctuate or even stagnate, whereas the first may close after an otherwise ordinary disappointment. A survival model can learn this difference without learning which venture will produce the next large jump in sales.
That distinction is especially important when moving the logic back to a business plan. A customer relationship may be evidence that a venture has begun converting an idea into an external exchange, and that exchange may contribute to resources or learning. But the relationship is not the same thing as a cash buffer, and a plan’s description is not the same thing as a ten-year account history. It is an early trace that may be useful precisely because it arrives before a complete track record exists. Its value should be judged by whether it adds information to the screening question, not by whether it recreates every later condition of survival.
The survival results move in the other direction. The survival regressions begin with a Nagelkerke R² of about 15% in year 2 and rise to roughly 26–36% by year 10. The percentage of cases correctly classified also increases in most years. The authors describe this as the survival fog clearing, although not through one dramatic step at a particular age. The visibility improves gradually and is not a promise that every later-stage venture is easy to evaluate.
Selection helps explain the pattern. Ventures that start with little resource buffer are more likely to be removed by early shocks. Those that remain may have accumulated resources or passed through a period in which their ability to continue has already been tested. The population of survivors is therefore not the same population as the original cohort. By year ten, a screener is not asking which of all new ventures will persist from the beginning; they are often looking at a more selected group with more information about who has endured.
This point matters for venture screening because a model can benefit from information that arrives only after the most fragile cases have revealed themselves. An early-stage business plan does not contain ten years of survival history. Customer relations can provide an earlier observable trace of market contact, but they do not recreate the resource buffer, the selection process, or the information gained through time. The companion study therefore supplies a boundary for the focal study rather than a validation of its customer coefficient.
The two studies also differ in their data and setting. The companion analysis uses detailed bank-account records for a large UK cohort that started trading in the same quarter. The customer-transaction study uses 272 technology-oriented business plans from a German regional setting, many submitted before legal incorporation, and asks about survival five years after submission. Their results should not be combined into one estimate. Their value lies in how they fit together conceptually: one explains why survival and growth can diverge as outcomes; the other tests a particular early textual signal against a survival outcome.
There are further reasons to resist a deterministic reading. The companion study acknowledges that some bank-account variables may be causes, consequences, or co-evolving indicators of performance. Its regression coefficients should therefore be read as partial associations rather than causal effects. Even the resource-threshold story is a useful model of selection, not proof that every venture exits for the same financial reason. A survival pattern can be statistically clearer without revealing a single mechanism that governs every firm.
This is why “the fog lifts” should be read carefully. It does not mean that time turns entrepreneurship into an engineering exercise. It means that, for the narrower binary question of whether a venture remains in operation, some signals become more informative as fragile firms exit and resources accumulate. The broader question—who will achieve unusually strong sales growth—can remain noisy throughout.
For the customer-transaction argument, the consequence is direct. A market relationship may be worth screening because it can serve as evidence that a venture has moved beyond intention toward exchange and resource formation. But if the measure is used to rank likely winners, it is being asked to answer a different question from the one the evidence supports. Survival can be a necessary condition for later achievement without being a sufficient indicator of it.
The next chapter turns from this outcome boundary to the signal itself. It examines what the study means by a transaction relation, how customer, financier, partner, and human-resource ties are staged, and why an observable relationship is treated as an indirect proxy for a venture’s ability to adapt rather than as a direct measure of future greatness.